Workstream 3: development of the casemix classification

The protocol for this study has been published55 was planned to meet objective 4: to develop a patient-centred palliative care casemix classification, based on individual patient and family needs, for adults with both cancer and non-cancer conditions in the last year of life.

Aim

Our aim was to develop a casemix classification for UK specialist palliative care, for use in hospital-based palliative care, inpatient hospices (palliative care units) and home-based palliative care.

Methods

Design

We designed a multicentre prospective cohort study, following patients during episodes of specialist palliative care and reported according to the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) statement.86

Definitions

We defined an episode of care as a ‘period of contact between a patient and palliative care service provider or team of providers that occurs in one setting’.87 A new referral into palliative care or change of setting (e.g. re-location into an inpatient setting from home/care home, or vice versa) signalled the start of an episode of palliative care. Discharge from palliative care, change of setting or death signalled the end of an episode of care. We defined the setting of care as one of three types: hospital advisory (hospital-based specialist palliative care teams providing an advisory or consultant service), inpatient hospice (where patients are admitted to a hospice or specialist palliative care unit for an overnight stay of one or more days) or community-based specialist palliative care (where the patient receives care in their usual place of residence, either at home or in a care home).

Population and settings

Patients were recruited from 14 organisations providing specialist palliative care services in England: four hospital advisory services, five inpatient hospice services and seven community-based services (some organisations provided more than one setting of care). Sites were selected to ensure our study sample was representative of UK palliative care patients and services in terms of participant age, ethnic background, service size, proportion of cancer/non-cancer patients and urban–rural balance. We included more community services than hospital or inpatient hospital services.

Inclusion/exclusion criteria

We included consecutive adult patients (≥ 18 years) receiving specialist palliative care at all participating sites. Exclusion criteria comprised patients aged < 18 years, those who declined participation and/or those who previously expressed a wish not to participate in research.

Data collection and primary outcome

Data were collected from clinicians between July 2015 and October 2016, with completion of follow-up at the end of January 2017; no data were collected directly from patients. Collected data included demographic and clinical variables, episode start and end dates, potential casemix variables, and data on patient-level and other costs of providing specialist palliative care. All participants received the usual specialist palliative care at that site, including a multidisciplinary team with specialist training delivering holistic care focused on physical and psychological symptom management, social/family support, planning ahead around priorities and preferences, and care into the dying phase including post-death care of the family, where relevant.

Potential casemix variables were selected based on (1) being patient-level attributes and (2) existing evidence of association with casemix/complexity.88 The key casemix variables included were age, sex, ethnicity, living circumstances, need for interpreter, primary diagnosis, palliative Phase of Illness, functional status, dependency and symptoms/problem severity.

Details of the measures used are presented in Appendix 1.

Our primary outcome was the cost of specialist palliative care per day. We adopted a (palliative) provider perspective for costs. The costs of acute hospital care, primary care, generic end-of-life care (i.e. provided by non-specialist teams), and informal care costs were excluded, not because these are unimportant, but because we sought casemix criteria relevant to specialist palliative care. A broader perspective on costs is planned for future work.

Palliative Phase of Illness was assessed daily for people receiving inpatient (hospital or inpatient unit) care and at every contact for those receiving community-based care. Each change in Phase of Illness (or end of episode) triggered the collection of the AKPS, IPOS or Palliative Care Problem Severity Score (PCPSS), and the Barthel Index. All staff involved in patient care recorded the time spent delivering care to participants at the patient level using the staff activity matrix.

We collected data from participating sites on the costs of delivering their services and patient-level resource use data from the staff activity matrix to derive actual patient-level costs according to a standard costing methodology based on current NHS costing principles.89 Note that costs captured for the hospital advisory and community-based settings represented the additional or ‘top-up’ costs for adding palliative care support to the hospital or community setting (and thus are reasonably compared). In contrast, the costs for inpatient hospices represented all the costs of inpatient care (and so are more reasonably compared with the costs of acute hospital admission). The full costing methodology (how costs were collected, classified and compiled) is available from the corresponding author on request.

Sample size

Based on standard recommendations for fitting multivariate models, a minimum of 50 + 8 × m cases for testing multiple correlation (where m is the number of predictors) are required to test the null hypothesis that the population multiple correlation equals zero with a power of 80%, α = 5% and a medium effect size for the regression analysis (R² = 0.13).90,91 The unit of analysis was episodes within sites; therefore, 10 predictors required 130 episodes per site. Allowing an additional 15% for episodes with missing data and 20% for cost outliers, we estimated that a target of 2674 episodes of care (191 episodes × 14 sites) was required.

Data handling

Data were collected prospectively at each site, recorded on an electronic database, cleaned and checked. Checked data were transferred to statistical software [Stata^® standard edition (SE) V.12 (StataCorp LP, College Station, TX, USA) and MATLAB^® 8.2 (The MathWorks, Inc., Natick, MA, USA)] for analysis.

Analysis

An exploratory data analysis was undertaken first, examining variables of interest one at a time. Descriptive statistics (means, standard deviations, medians, ranges and correlations) were calculated. Comparative box plots were constructed to investigate the differences between sites, episodes and phases. Cost of care was used as the response variable, measured as the total cost per episode.

The aim of the analysis was to form distinct groups within the data, such that patients within each group were similar to each other, but different from patients in the other groups.

To discover which baseline casemix variables could best predict the cost of a particular episode of care, the following nine steps were then undertaken:

• We removed incomplete episodes, retaining only the complete episodes of care.
• Following a previously adopted approach,46 high- and low-cost outliers were identified and removed using a trimming algorithm based on the IQR with the upper trim point at Q3 + 1.5 IQR and the lower trim point at Q1 – 1.5 IQR (where Q1 is first quartile and Q3 is third quartile).46 The trimming algorithm was applied to each setting separately.
• We examined the distribution of costs of specialist palliative care, by setting.
• Then, following the same approach as the development of the Australian casemix classification,88 we developed and validated a cost-predictive model using classification and regression tree (CART) analysis, which constructs decision rules in a hierarchical manner to form a branching classification.
• We used CART analysis to enable the more complex interactions between the predictor variables (both categorical and continuous) to be explored. CART has the advantage of being non-parametric and is not significantly impacted by outliers in the input variables. It enables the use of each variable more than once, if required for the optimal regression tree.
• Explanatory variables were compared to find the one which could best split the data into two homogeneous groups that were as different from one another as possible. These two groups would then be further split, using the same or another explanatory variable. Successive binary splits were performed on the data until there was no further improvement to be made and the best possible classification solution was reached.
• The best CART was deemed to be that which accounted for the largest proportion of variation in the cost of care (the response variable). The criterion used to compare the different ‘trees’ was the proportion of the variance of the response variable that could be explained by the selected groups.
• Costs were log-transformed for better modelling and back-transformed for providing mean costs per class or at each terminal node. Decisions about rules for splitting were informed by clinical utility (for instance, allowing branches that made clinical – as well as statistical – sense), as well as statistical performance as outlined in 7. We selected a maximum of four branches and a minimum of 30 cases per branch, for reasons of clinical utility. 10–fold cross-validation was used to prevent overfitting of the developed classification. No recalibration was undertaken. The analysis was done in R 3.5 (The R Foundation for Statistical Computing, Vienna, Austria) using the rpart, caret and (for bagging the trees) RWeka packages.
• For each setting, we reported the variance explained by the CART model and the root-mean-squared error (RMSE) (i.e. the square root of the variance of the residuals). The RMSE indicates the absolute (rather than relative, as with R²) fit of the model to the data.

Ethics approval

The trial registration number is ISRCTN90752212. Ethics approval was received from the Camberwell St Giles National Research Ethics Service REC on 2 July 2015 [REC Reference: 15/LO/0887, Integrated Research Application System (IRAS) Project ID:172938].

Results

Subject characteristics

A total of 2469 patients were recruited, providing data on 2968 complete episodes of specialist palliative care (12 incomplete episodes were removed prior to analysis); 2087 participants contributed one episode of care, 283 participants contributed two episodes and 99 contributed three or more episodes. Demographic and clinical characteristics for the 2469 participants are reported in Table 1. No participants withdrew after recruitment. Further details are available in Appendix 2.

Episode characteristics

Details of the 2968 complete episodes of care and related casemix variables are reported in Table 2.

Outliers

Using the trimming algorithm described in Methods, Analysis, 123 (16.0%) hospital advisory episodes, 185 (24.2%) inpatient hospice episodes and 305 (21.2%) community episodes were removed to ensure that the principal cost and classification reporting was not based on outliers (a common challenge in costing studies).

Costs of specialist palliative care

The distribution of the total cost of specialist palliative care episodes, derived from the trimmed data set, is shown in Table 3; costs are shown (1) by day, (2) by day, broken down by Phase of Illness, and (3) by episode of care, together with details of length of episodes.

Classification and regression tree analysis

Figures 4–6 show the CARTs for each setting. The per cent variance explained (and RMSE) were 20% (RMSE = 0.30), 51% (RMSE = 0.51) and 27% (RMSE = 0.36), for hospital advisory, inpatient hospice and community episodes, respectively.

FIGURE 4 Classification tree of casemix criteria for hospital advisory episodes of specialist palliative care (costs per day reported for each class). (Figure not reproduced here; see the original article.)

FIGURE 5 Classification tree of casemix criteria for inpatient hospice episodes of specialist palliative care (costs per day reported for each class). (Figure not reproduced here; see the original article.)

FIGURE 6 Classification tree of casemix criteria for community episodes of specialist palliative care (costs per day reported for each class). (Figure not reproduced here; see the original article.)

Seven different casemix variables provide the optimal combination to develop classes for each of the settings. Table 4 shows which variables were used and how they were combined to constitute the casemix classes, including cost weights.